A Windows PC
With a recent NVIDIA graphics card, Pollen's tools run in Windows' own Linux and train on your card.
A Linux PC
With a recent NVIDIA graphics card, Pollen's tools run as they come and train on your card.
A Mac
With Apple silicon, you come in through Microduck Studio for Mac or Pollen's workshop, and train on a card you rent.
An iPhone or iPad
Microduck Studio compares, tunes and searches on the device, and trains with Pollen's trainer on a card you rent.
A PC with no NVIDIA card, or an older one, still follows its own guide, and the guide says where to switch to a rented card.
Why some machines rent a card
Why can a Mac or an iPhone not run Pollen's trainer itself? Pollen Robotics train the Microduck with mjlab, which needs an NVIDIA graphics card to train: its README says macOS is supported for evaluation only, and mjlab is Python, which an iPhone cannot run.
So those machines rent the card from Hugging Face Jobs by the minute, on your own Hugging Face account, which needs credit on it. The card Pollen's workshop and Microduck Studio both choose, one NVIDIA L4, costs $0.80 an hour at Hugging Face's prices.
What runs where
Everything around the training happens on your machine: choosing what to train, judging what comes back, watching it walk and publishing it. Only the training itself needs the graphics card, so only the machines without a suitable one send it away.
Microduck Studio works the same way. It is our app for iPhone, iPad and Mac, and it starts with what the device can do on its own. Because MuJoCo runs on the device, it compares two networks, tunes one and searches all of a network's weights right there, with nothing sent anywhere.
It goes online only when you ask it to: to fetch a network or publish one, to rent a card at Hugging Face, to use a bench on your own network for faster physics, or to draft a motion from words with a model on another computer of yours or a service you have a key for. Drafting can stay on the device too, with Apple's on-device model where yours has one.
Whose tools these are. The trainer, the challenges and the Arena are Pollen Robotics', built on mjlab and MuJoCo. Microduck Studio is an independent project, and so is the recipe it rents a card for, which fine-tunes Pollen's kick, or a smaller network distilled from Pollen's walk, with Pollen's own trainer. If a guide goes wrong for you, tell us which step.
Then
While it trains, the books explain what is happening: the reward, the practice worlds and the trainer, with the real numbers.
After your first duck, Go further reads a real training run from its records, runs one change of your own on a rented card, and reaches the duck's other channels, from its camera to its spare Qwiic port.
When you have a policy, Race in Pollen's Arena says what each of the Arena's events really tests, which lessons explain it, and what our own policies did under its rules.